We analyse the strategies used in formulating situated informal location descriptions, by identifying geospatial expressions contained therein and annotating each for properties such as geospatial granularity and identi ability. Analysis of the annotations leads to insights such as the predominance of suburb-level expressions, and prevalence of vernacular expressions.
When informally describing one's whereabouts or giving directions, people make heavy use of place descriptions. In the descriptions they relate their location to the surrounding objects, or landmarks [1]. In order to make the instructions interpretable by the recipient, the description provider should use familiar landmarks and relate the location to them appropriately. Thus, for a human recipient this task is trivial. However, computational systems cannot easily interpret place descriptions expressed in natural language, or generate natural-sounding route or place descriptions.
Additionally, humans frequently make use of vernacular place descriptions, or refer to landmarks using non-standard renderings of their `o cial' names, as a result, making it hard for computers to understand the description, and also humans unfamiliar with the locality being described. Wu and Winter state that placenames and spatial relations are main components of place descriptions, and in order to interpret the descriptions their components must be interpretable [2].
In this study we focus on analyzing placenames in the context of informal place descriptions, that is placenames that are elicited naturally and in situ, without any constraints or guidance. We manually identify geospatial expressions in a dataset of placename descriptions, and further annotate the granularity level, identi ability and normalised name of each such expression. 2
Winter et al. collected situated place descriptions from players of the Tell us where location-based mobile game [3].1 The game consisted of submitting textual descriptions of the location of smart phone users, along with their GPS location. The reasons we chose to use this data are many fold. First, the data was collected across a broad sample of users, ensuring the heterogeneity of the data and reducing sample bias. Second, the participants were asked to submit textual descriptions of their location from anywhere in the state of Victoria, Australia. This led to a diversity of locations, but within a restricted area of familiarity to our annotators and with the expectation of consistency in the strategies used by the participants to describe their location. Third, the users were given no guidelines for writing the descriptions, meaning that the data is rich in vernacular placename descriptions and the strategies used by users to describe their location are varied. Lastly, since the participants were using their mobile phones and basing the placename descriptions on their actual location. As a result, the descriptions are situated, spontaneous, and as natural as we could hope for.
A total of 2221 place descriptions were collected through the Tell us where game. However, the data contained duplicates. Since we are interested in qualitative rather than quantitative data, it was decided to eliminate all duplicates from the corpus. As a result, the nal number of descriptions was 1858. 2.1
Annotation
We manually annotated the placename descriptions for geospatial expressions, in
the form of: (
One of the broader goals of this work is the compositional semantic interpretation of place descriptions. It was thus decided that we should aim for maximum segmentation granularity in our annotation, while avoiding nested annotations. For example, if the place description were an address such as Melbourne University Bookshop, in Parkville near the library, we would segment it into the geospatial named entities Melbourne University Bookshop and Parkville, and the geospatial noun phrase the library. Note that we would not also identify Melbourne University as a geospatial named entity, as it is nested within Melbourne University Bookshop.
We expected many of the geospatial expressions in the dataset to be noun chunks. For example, Queen Victoria Market is a single noun chunk geospatial named entity, while a tall building is a single noun chunk geospatial noun phrase
Granularity level Description
(
Hwy), islands (French Island ), rivers (Murray river ) and states (WA) Table 1. Granularity level classi cation (Richter et al., 2012); all examples are taken from the actual dataset, and are presented using the original orthography referring to a construction, which can be used as a reference point when describing a location. In the interests of expediting annotation, we rst chunk-parsed the place descriptions, using the Stanford CoreNLP tools.
The annotation scheme we used is comprised of several layers. The rst annotation layer contains information about whether a given segment is a geospatial named entity (NE NP ) or a geospatial noun phrase (NP NP ). The remaining layers apply to each geospatial expression.
The second layer of annotation is the granularity level, and captures the \zoom level" of each geospatial expression. The granularity level is judged on the scale from 1 to 7, based on the classi cation of Richter[4] as detailed in Table 1. In some instances, we diverge from Richter's classi cation. For example, when a named entity is too big or too small for the bounding box of its default zoom level, we override the default to capture the zoom level which best matches the size of the bounding box. Mountain Highway, e.g., goes through only a few suburbs of Melbourne, so we override the Country granularity level for highways and assign it to the zoom level of City to better re ect its size. Similarly, when determining the granularity level of towns, it was decided to shift the small towns that do not have suburbs (e.g. Warragul and Pakenham) from City to District.
The third layer of annotation is identi ability. This captures whether a geospatial expression is unique within Victoria or there are multiple instances of it. There are three possible values for identi ability: non-identi able, identiable ambiguous, and identi able non-ambiguous. All geospatial noun phrases (e.g. school, park, monument ) are non-identi able, since the set of these objects within Victoria is very large and it is not possible to geocode them without disambiguating information. Some geospatial named entities are considered to be non-identi able due to their ubiquity and unavailability within standard gazetteers of an exhaustive listing of every instance within Victoria (e.g. McDonalds, 7-eleven). On the other hand, a geospatial named entity can refer to a small set of several places which are enumerated in a gazetteer, in which cause they are considered to be identi able ambiguous. For example, there are four instances of Canning Street in Victoria, so every Canning Street in the corpus is annotated as identi able and ambiguous. On the other hand, Flemington Road is identi able non-ambiguous as there is only one instance in Victoria.
As with granularity, the determination of identi ability is inevitably subjective. To reduce the e ects of subjectivity as much as possible, we base the judgement on two online gazetteers: OpenStreetMap2 and Google Maps.3. Google Maps contains an extensive listing of named entities, but has poor coverage over non-standard or vernacular equivalents of less well-known named entities. Thus, while melb uni (standard = The University of Melbourne) and fed square (standard = Federation Square) can be found in Google Maps, it does not contain local vernacular such as broady (standard = Broadmeadows ) or non-standard abbreviations such as pi for Phillip Island or fg for Ferntree Gully. Here, we elicited support from locals and the Google search engine to interpret the geospatial expression.
Names of cafes, restaurants, and other small businesses were the most di cult NEs to judge identi ability for. Even though OpenStreetMap lists a vast number of buildings, eating places, shops, many of them were missing.
The fourth and nal level of annotation is the placename normalisation. Since the place descriptions were submitted by mobile phone, the dataset contains a lot of abbreviations, misspellings and vernacular names. The canonical name/spelling was provided in all such instances. For example, melb uni would be normalised to The University of Melbourne. We observed an inevitable dependency between identi ability and placename normalisation for geospatial named entities: if a geospatial named entity cannot be identi ed, it is not possible to determine its normalised rendering.
Some of the submitted place descriptions do not contain any information about the location (e.g. this will be an everlasting love) or are located outside of Victoria (e.g. in Wagga Wagga). All such descriptions were marked as irrelevant at the message level, using the IRREL label.
For the annotation we used brat,4 a highly-con gurable, easy-to-use webbased text annotation tool. 3
Having annotated the dataset, we extracted a feature vector for every annotated geospatial expression (excluding the irrelevant descriptions). Each feature vector contained a set of values: id, geospatial expression type, granularity level, identi ability, original spelling, and canonic (normalized) spelling. Then, all the vectors
were collated into a table and fed into the R statistical package5 for analysis. In total, 3061 geospatial expressions were extracted, 2139 (70%) of which were geospatial named entities. That is, without any constraint on the description, about two thirds of geospatial expressions contained in place descriptions can potentially be found in gazetteers.
Figure 1 presents a distribution of geospatial expressions across zoom levels, broken down by identi ability. The mean granularity value is 4.01, with a standard deviation of 1.05. The most common granularity level is 4 (Street ), with about 45% of all geospatial expressions. This means that when writing place descriptions, users tend to make heavy use of streets, parks, squares, universities and hospitals. Of the remainder, almost a quarter (24%) of the referents are of the Building granularity level (level 3), and about 18% are of the (Suburb) granularity level (level 5).
The correlation between the granularity level and the fraction of non-identi able placenames is not very surprising: the bigger the spatial feature, the more likely it will be identi able. On the other hand, the appreciable drop in non-identi ability at the Suburb level is proof of the salience and unambiguity of the placenames within this level. After dividing all the geospatial expressions by identi ability and ltering out from the non-identi able ones the names of chain stores and eating places (e.g., McDonald's, Subway, Coles ), it is possible to calculate how many of the named entities are not in the gazetteers (Open
StreetMap and Google Maps). Out of 2139 named entities, 51 (2.4%) are not contained in the gazetteers. As a rule, among these placenames are names of restaurants, apartment blocks, and other small scale companies (e.g. Pilkington Glass, Ching Chong Food, Yarra Crest Appartments ).
Another important category of geospatial expression is vernacular descriptions. We found a considerable number of entrenched vernacular equivalents of salient Victorian placenames, and common strategies for forming vernacular place names. Some of them are formed by simply dropping one of the constituent words (Narre Warren ! narre), some by \clipping" the word (Yackandandah ! yack, Dandenong ! dande), and some are acronyms (Phillip Island ! pi, Ferntree Gully ! fg ). However, the most productive pattern was \embellished clipping", shortening the expression to the rst syllable and adding a diminutive su x -y, -ie, (e.g. Richmond ! richy, Beacons eld ! beacy, South Gippsland Highway ! south gippy ). The pattern is particularly peculiar to the Australian English. From the collected informal NEs, one can infer that only salient and unambiguous placenames undergo the process of vernacularization. Since suburb names in Victoria are unique and widely used for describing locations, they are most commonly substituted by their informal equivalents. 4
In this paper, we have performed detailed component-wise analysis of informal place descriptions. From this study, we can conclude the following: (a) most geospatial expressions are streetnames, parks, buildings and suburbs; (b) the presence of a suburb-level placename in the description increases its identi ability; (c) vernacular place descriptions are commonly used, based on a small number of strategies; and (d) geospatial named entities which are mostly likely to not be contained in gazetteers are names of pubs, cafes, and small businesses.
This paper has considered placenames independently of the message-level interpretation. A logical next step is a compositional analysis of the place description based on the annotations we have done, and investigation of how spatial relational semantics (e.g, prepositions like near, at, in) impacts on message interpretability and the properties of its constituent placenames.